Final release-readiness verification¶

Verified locally on 2026-10-13 for version 0.1.0. Recommendation: NOT READY for public release until the configured GitHub Actions workflow passes, including Windows and macOS. All executed Linux checks passed. No new release-blocking implementation defect was reproduced in this final pass. This report records finite benchmark evidence, not a guarantee for arbitrary models, callbacks, or numerical platforms.

1. Feature-completion matrix¶

2. Public API and architecture¶

There are exactly two functional submodules: curvefit_stats.inference and curvefit_stats.estimators. __init__.py re-exports the same ten objects without wrappers. Direct module imports remain available. There are no additional functional modules in the source or wheel. Typed result dataclasses remain inside estimators.py; py.typed is included in both distribution types.

The signature/docstring consistency tests passed for all ten root exports and GlobalFitGroup.fit. No signature or runtime source changed in this final pass. The exact signatures, annotations, NumPy-style Parameters/Returns/Raises/Notes/ Examples sections and return shapes are in the API reference.

Bands return lower/upper arrays; audit and collinearity return metric dictionaries; summary returns a table string; profiles return per-parameter error dictionaries; bootstrap returns mean parameters, empirical covariance and successful samples; bounded covariance returns a parameter covariance matrix; propagation returns sample summaries. GlobalFitGroup(...).fit(...) returns a typed joint-fit result. EIV returns (popt, pcov), or a third typed diagnostic result with full_output=True.

Runtime source import inspection and built metadata confirm only the standard library, NumPy and SciPy. Ruff found no unnecessary imports. Source inspection found no unfinished implementation, debug statements or duplicate public implementations. Ellipsis-only bodies are intentional typing overloads for EIV.

3. Tests, coverage and quality checks¶

The hash-locked Linux Python 3.12 development environment passed 405 tests with 95.27% combined statement/branch coverage, exceeding the 90% gate. The five installed-wheel suites below each passed all 405 tests, with no skips or expected failures. All use strict pytest configuration and warning handling.

• Ruff lint passed; formatting passed for all 15 Python files.
• Strict mypy passed for all 15 source/test/example/tool files.
• Thirteen doctests passed in the source environment and each wheel environment. One explicitly expected bound-active conditional-covariance warning was emitted.
• All eleven executable workflows passed, as did README, user-guide and integration Python snippets. Tests distinguish narrowly expected warnings from failures.
• Actionlint 1.7.7 and git diff --check passed. This validates workflow syntax; it does not establish successful execution on GitHub's runners.
• Isolated PEP 517 sdist and wheel builds passed. Strict Twine validation and archive-content/metadata checks passed. The wheel was built from the sdist.
• Isolated installed-package imports and all-ten-feature smoke checks passed in each Linux wheel environment and separate runtime-only wheel/sdist environments. Direct sdist installation used an isolated backend. pip check passed in all eight checked environments. Runtime-only distributions were exactly pip, curvefit-stats, NumPy and SciPy, with no development tools.
• Two independent fixed-epoch wheel builds were byte-identical. This result is specific to this backend/platform/epoch; sdist gzip or cross-platform numerical byte reproducibility is not asserted.

4. Numerical validation¶

Executed independent references include linear regression, exponential decay, Michaelis–Menten kinetics, logistic growth and correlated nonlinear models. The stress/regression suites exercise heteroscedastic errors, very small/large uncertainties and response units, large parameter-unit differences, nearly singular Jacobians, active constraints and large floating-point offsets. Pathological uncertainties, covariance matrices, observation shapes, insufficient dof and failed optimizations are separately tested. Stochastic comparisons use controlled local generators and sampling tolerances; repeatability does not imply identical optimizer/LAPACK results across platforms.

Covariance and identifiability claims were checked against noise provenance: shared parameters occupy one joint optimization column, conditional bound covariance leaves active rows/columns NaN, and EIV covariance does not turn response-only likelihood methods into ODR methods. Failed optimization and unidentified models do not receive fabricated usable uncertainty. See the scientific review for the nine previously corrected defects and their regression cases; all 46 review regression cases passed again.

5. Executed package compatibility matrix¶

Metadata specifies Python >=3.11, NumPy >=1.26.4,<3 and SciPy >=1.12,<1.19. These ranges are compatibility policy, not exhaustive testing of every version combination. The reference development extra/lock uses Python 3.12 because of the official SciPy stubs; runtime/test-extra compatibility includes Python 3.11.

6. Remaining documented scientific limitations¶

• Bands are pointwise; nonlinear Delta/Wald inference is local and approximate. fit_summary always uses Student-t with N-p; known-error normal inference is available for bands, not that table. Covariance arrays cannot encode provenance.
• Regular AIC/BIC assume maximum-likelihood estimates and valid parameter counts; rank screening does not establish optimality, structural identification or boundary regularity. Profile thresholds use asymptotic Wilks calibration and local nuisance optimization; boundaries can require separate calibration.
• Residual bootstrap assumes IID observation errors at fixed predictors; fitted residuals are leverage-dependent. It provides neither an unbiased finite-sample covariance guarantee nor automatic corrections for leverage or failed fits.
• Bound-active covariance holds selected parameters fixed. It is not unrestricted marginal covariance, active-parameter uncertainty or selection-calibrated inference.
• Gaussian propagation may violate bounds; rejecting derived outputs changes the summarized distribution. Finite ratio samples do not imply finite ratio moments.
• Global fitting estimates one common unknown noise multiplier; separate unknown dataset scales and cross-dataset error covariance are not modeled.
• EIV supports one predictor/response and scalar/vector axis uncertainties, without bounds, correlated axis errors or analytic ODR Jacobians. SciPy ODR is deprecated from 1.17 and removed in 1.19; calls emit an explicit compatibility warning on deprecated versions and the dependency upper bound excludes removal.
• Float64 range, numerical rank thresholds and model scaling limit arbitrary callbacks. Numerical convergence does not establish a global optimum.

These are scientific/API limits, distinct from reproduced software defects. They are explained in the user guide and integration guidance.

7. Release blockers and 8. recommendation¶

NOT READY: the candidate's actual GitHub Actions workflow, including its Windows and macOS matrix jobs, has not executed. Local Linux equivalents and Actionlint are insufficient evidence for that platform release gate. No further release-blocking code defect was reproduced. The next checkpoint is successful execution of the existing workflow on the candidate revision before public release. No push, workflow dispatch or publication was performed.

Only this report was added in the final verification phase; runtime APIs and implementations were preserved. Distribution artifacts were regenerated locally.

Executed commands¶

Commands used the repository cwd /workspace/curvefit-stats, except isolated smokes and dependency checks, which used /tmp. Logs and coverage XML files are under /workspace/setup-checks/curvefit-stats.

.venv/bin/ruff check .
.venv/bin/ruff format --check .
.venv/bin/mypy
COVERAGE_FILE=/workspace/setup-checks/curvefit-stats/.coverage-release /workspace/setup-checks/curvefit-stats/locked-dev/bin/python -m pytest -q --cov-report=xml:/workspace/setup-checks/curvefit-stats/coverage-release.xml
.venv/bin/python -m pytest -q --no-cov --doctest-modules src/curvefit_stats examples -W 'default:Bound-active parameter rows:UserWarning'
.venv/bin/python examples/workflows.py
SOURCE_DATE_EPOCH=1767225600 PIP_CACHE_DIR=/workspace/.cache/pip .venv/bin/python -m build
SOURCE_DATE_EPOCH=1767225600 PIP_CACHE_DIR=/workspace/.cache/pip .venv/bin/python -m build --wheel --outdir /workspace/setup-checks/curvefit-stats/release-repeat-dist
.venv/bin/python -m twine check --strict dist/*
.venv/bin/python tools/check_distributions.py
/workspace/setup-checks/curvefit-stats/actionlint/actionlint
git diff --check

For each ENV in py311, compat-venv, wheel312, py313, and py314, PYTHON was /workspace/setup-checks/curvefit-stats/ENV/bin/python. Each ran:

UV_CACHE_DIR=/workspace/.cache/uv uv pip install --python PYTHON --reinstall --no-deps dist/curvefit_stats-0.1.0-py3-none-any.whl
COVERAGE_FILE=/workspace/setup-checks/curvefit-stats/.coverage-release-ENV PYTHON -m pytest -q --cov-report=xml:/workspace/setup-checks/curvefit-stats/coverage-release-ENV.xml
PYTHON -m pytest -q -c pyproject.toml --no-cov --doctest-modules INSTALLED_PACKAGE_PATH examples -W 'default:Bound-active parameter rows:UserWarning'
PYTHON -I /workspace/curvefit-stats/tools/check_install.py
PYTHON -m pip check

The installed package path was obtained from that interpreter's curvefit_stats.__file__, never by setting PYTHONPATH=src for artifact checks. The separate runtime-only checks ran:

UV_CACHE_DIR=/workspace/.cache/uv uv pip install --python /workspace/setup-checks/curvefit-stats/wheel-runtime/bin/python --reinstall --no-deps dist/curvefit_stats-0.1.0-py3-none-any.whl
PIP_CACHE_DIR=/workspace/.cache/pip /workspace/setup-checks/curvefit-stats/sdist312/bin/python -m pip install --no-cache-dir --force-reinstall --no-deps dist/curvefit_stats-0.1.0.tar.gz
/workspace/setup-checks/curvefit-stats/wheel-runtime/bin/python -I /workspace/curvefit-stats/tools/check_install.py
/workspace/setup-checks/curvefit-stats/sdist312/bin/python -I /workspace/curvefit-stats/tools/check_install.py

Both runtime interpreters and locked-dev also passed python -m pip check. Additional stdlib probes verified module enumeration, signatures, AST import roots/intentional overload declarations, runtime distribution inventory, artifact source equality and fixed-epoch wheel-byte equality.